Multi-modal Face Recognition

Transcription

1 Multi-modal Face Recognition Hu Han /04/06

2 Outline Background Related work Multi-modal & cross-modal FR Trend on multi-modal (face) recognition Conclusion and discussion 2

3 Background Unconstrained sensing & uncooperative subject scenario poses great challenges to unimodal FR system It s our intention to go through every frame of every video. Boston Police Commissioner, Ed Davis We are particularly interested in reviewing video footage captured by bystanders with cell phones or personal cameras near either of the blasts In an investigation of this nature, no detail is too small. Attorney General, Eric H. Holder Jr. 3

4 Background Unimodal FR A manually selected probe face image of the suspect (Tamerlan Tsarnaev) with the best quality is matched with its true mate by a COTS with rank-5000 among a 1M gallery set Probe 1M 4

5 Background Challenges Low quality surveillance videos and images Intentional thwarting of identification (e.g. sunglasses and hats) Daunting amount of data Videos or images are n/a Multi-modal FR is a possible solution Advances in computing and imaging tech. RGB, depth, NIR, 3D, sketch, etc. Multi-modality, multi-view, multi-biometrics 5

6 Background Top K Matche s Top K Matche s Human operators manually review K*n images (n = # of images in the face media collection) Traditional Forensic Investigation Workflow Top K 2016/4/6 Matche hanhu@ict.ac.cn 6 s

7 Outline Background Related work Multi-modal & cross-modal FR Trend on multi-modal (face) recognition Conclusion and discussion 7

8 Related work Multi-modal FR 2D + 3D Beumier and Acheroy, PRL 01 Chang et al., ACM-W 03 2D + depth Lu and Jain, TPAMI 06 2D + 3D + NIR Bowyer et al., Most are: Per-modal matching + score-level fusion 8

9 Related work Cross-modal FR Modality transformation Wang & Tang, TPAMI 09 (sketch vs. photo) Gao et al., TCSVT 12 (sketch vs. photo) 3D face modeling, Blanz & Vetter 03 (2D vs. 3D) Invariant features Lei & Li, CVPR 09 VIS-NIR Klare & Jain, TPAMI 13; Han & Jain TIFS 13; Klum et al., TIFS 14 VIS-NIR, forensic sketch, VIS-TIR 9

10 Outline Background Related work Multi-modal & cross-modal FR Multi-modal FR Trend on multi-modal (face) recognition Conclusion and discussion 10

11 ... Multi-modal face recognition Still image Sketch 11? Video 3D Human operators manually review 1*K images

12 Multi-modal face recognition A hierarchical quality-based fusion 岁男性白人 Image/video Sketch 3D Quality measures: Q1 Q2 Q3 Q4 1M Mugshot,True mate is matched at rank-112 (vs. rank-5000 in unimodal) 12

13 Multi-modal face recognition Face Track Extraction U u 1,u 2,...,u a... All Frame Pairs... V v 1,v 2,...,v b Same Not Same t t su,v Matching a Face Track from a Video Multiframe Score level Fusion: mean median max min su 1,v Similarity Matrix COTS Face Matcher... su a,v b a b 13

14 Multi-modal face recognition Pose Correction via 3D Face Modeling 14

15 Multi-modal face recognition Close set identification 4,249 gallery images, 596 probe subjects Images Videos Multimodal 15

16 Multi-modal face recognition Close set identification 4,249 gallery images + 1M background mugshots, 596 probe subjects 16

17 Multi-modal face recognition Open set identification The person of interest may not be present in legacy face databases The gallery consists of 596 subjects with at least two images in the LFW database and at least one video in the YTF database 17

18 Multi-modal face recognition Quality based fusion 18

19 Multi-modal face recognition A Case Study on the Boston Bomber (Gallery of one million mugshots) 19

20 Multi-modal face recognition Forensic Sketches from Low Quality Video Retrieval ranks without and with 2016/4/6 demographic filtering are given as #(#) 20

21 Deep multi-modal FR CNNs CNNs CNNs CNNs CNNs CNNs RGB D RGB D 21

22 Deep multi-modal FR Deep RGBD face recognition 900,000 RGBD images of 700 subjects Modality Accuracy RGB 0.93 Depth 0.86 Deep RGB-D

23 Outline Background Related work Multi-modal & cross-modal FR Cross-modal FR Trend on multi-modal (face) recognition Conclusion and discussion 23

24 Cross-modal face recognition Compatible with huge existing 2D face images RGBD vs. RGB Modality is not available NIR vs. VIS Sketch vs. photo 24

25 Cross-modal face recognition Sketch to mugshot matching Viewed sketch: drawing/synthesizing a sketch while looking at a subject/photo Forensic sketch: drawing/synthesizing a sketch based on verbal description from the victim or eyewitness COTS matcher for photo-to-photo matching can achieve over than 85% rank-1 identification rate for viewed sketch in 2013, while its performance for forensic sketch identification is less than 10% 25

26 Cross-modal face recognition Sketch to mugshot matching 26

27 Sketch to mugshot matching Sketch leads to arrest of suspects Timothy McVeigh (the David Berkowitz Oklahoma City bomber) (Son of Sam) Ted Kaczynski (the Unabomber) 27

28 Sketch to mugshot matching Local and holistic matching Local matching: component based rep. Hu Han, Brendan Klare, Kathryn Bonnen, and Anil K. Jain. Matching Composite Sketches to Face Photos: A Component Based Approach. IEEE Transactions on Information Forensics and Security (T-IFS), vol. 8, no. 1, pp , Jan

29 Sketch to mugshot matching Component based rep. is an inverse process of sketch composition 29

30 Sketch to mugshot matching Local and holistic matching Holistic matching: dense keypoint features 30

31 Sketch to mugshot matching Complementarity 31

32 Sketch to mugshot matching Hand-drawn and software-generated forensic sketch 32

33 Sketch to mugshot matching Software-generated viewed sketch 33

34 Generalized cross-modal FR Cross-distance and cross-spectral matching in nighttime FR 150m NIR at night Enrolled VIS 34

35 Cross-distance and crossspectral FR A learning based image restoration method to recover a high-quality face image from a low-quality Dictionary building 35

36 Cross-distance and crossspectral FR Per-patch recovery using LLE 36

37 Cross-distance and crossspectral FR Restored images 37

38 Cross-distance and crossspectral FR Cross-distance and intra-spectral test 38

39 Cross-distance and crossspectral FR Cross-distance and cross-spectral test 39

40 Sketch to mugshot matching Licensed to MorphoTrak, one of the world s leading biometrics companies Nighttime FR system received funding from FBI 40

41 Outline Background Related work Multi-modal & cross-modal FR Multi-modal FR Trend on multi-modal (face) recognition Conclusion and discussion 41

42 Trend on multi-modal (face) recognition The other biometrics or multi-biometrics Tattoo, gesture, Google Abacus project General object A. Wang, J. Lu, J. Cai, T.-J. Cham, and G. Wang. Large-Margin Multi-Modal Deep Learning for RGB-D Object Recognition, IEEE Trans. Multimedia, 17(11): , Nov RGB-D, 300 common household objects 42

43 Trend on multi-modal (face) recognition The other biometrics or multi-biometrics Tattoo, gesture, Google Abacus project Masked ringleader of crowd trouble during Italy-Serbia clash identified by his tattooed arms [1]. [1] [2] Hu Han and Anil K. Jain. Tattoo Based Identification: Sketch to Image Matching. ICB,

44 Trend on multi-modal (face) recognition The other biometrics or multi-biometrics Tattoo, gesture, Google Abacus Project (Google I/O 2015) 44

45 Trend on multi-modal (face) recognition The other biometrics or multi-biometrics Tattoo, gesture, Google Abacus Project (Google I/O 2015) 3TB data You are your password! 45

46 Outline Background Related work Multi-modal & cross-modal FR Multi-modal FR Trend on multi-modal (face) recognition Conclusion and discussion 46

47 What I want to convey... Multi-modal FR significantly boosts the face recognition performance, particularly in unconstrained scenarios; but the optimum process pipelines of individual modalities and fusing scheme are still not known Cross-modality FR, particularly forensic sketch recognition, has wide applications, but remains an open problem Download [Data] Still & video & sketch & 3D face images [Data] Cross-distance, cross-spectral face images [Data] Computer generated viewed-sketches [Protocol] Open-set identification protocol 47

48 Related papers [1] Lacey Best-Rowden, Hu Han*, Charles Otto, Brendan Klare, and Anil K. Jain. Unconstrained Face Recognition: Identifying a Person of Interest from a Media Collection, IEEE Transactions on Information Forensics and Security (T-IFS), vol. 9, no. 12, pp , Dec [2] Scott Klum, Hu Han*, Brendan Klare, and Anil K. Jain. The FaceSketchID System: Matching Facial Composites to Mugshots. IEEE Transactions on Information Forensics and Security (T-IFS), vol. 9, no. 12, pp , Dec [3] Dongoh Kang, Hu Han, Anil K. Jain, and Seong-Whan Lee. Nighttime Face Recognition at Large Standoff: Cross-Distance and Cross-Spectral Matching, Pattern Recognition (P.R.), vol. 47, no. 12, pp , Dec [4] Hu Han, Brendan Klare, Kathryn Bonnen, and Anil K. Jain. Matching Composite Sketches to Face Photos: A Component Based Approach. IEEE Transactions on Information Forensics and Security (T-IFS), vol. 8, no. 1, pp , Jan [5] Hu Han and Anil K. Jain. Tattoo Based Identification: Sketch to Image Matching. In Proc. 6th IAPR International Conference on Biometrics (ICB), Madrid, Spain, June 4-7, (Oral) 48

49 Thank You! 49